Agister: Agent Work as Settings, Not Code

The projects earlier in this series — RexLab and FolioTell — were built with AI coding agents. After a few months of that, the hard part was no longer getting an agent to write code. It was everything around it: which model with which effort for which job, what each call was allowed to touch, how a review was put together from several lenses, and, afterwards, what actually happened in a run that went wrong.

I kept solving these with shell scripts and prompt files, and kept losing track. Agister is the tool that replaced them. The code is at github.com/igor1309/agister.

Three kinds of settings

Agister describes work in three kinds of documents:

• An agent picks a runtime — Claude Code, Codex or bash — and, for a model, which model and effort.
• A process says what one call needs: its inputs, its prompt, the shape of its output as a JSON schema, and the permissions it asks for.
• A workflow composes steps — each a process on an agent — and other workflows: in sequence, in parallel, on a branch, or in a bounded repeat.

That separation is the point. "Review this diff for test quality" is a process. "Use a strong model at high effort" is an agent. "Run five review lenses in parallel, then have a judge merge them" is a workflow. Each can change without touching the others, and nothing about the work lives in Go code. Code adds only a new runtime or a new kind of input source.

Resolve first, then run

Before anything runs, Agister resolves the request into definitions or a refusal. A process that needs a permission its agent's runtime cannot grant is refused. A workflow that references a missing document is refused, with the chain that leads to the fault. agister check resolves every document and refuses exactly what a run would.

That gate matters with agents. A misconfigured run that starts, spends twenty minutes and money, and fails at step four is far worse than one that never starts.

Every call ends the same way

Whatever the runtime, every call ends in one result shape: status, output checked against the process's schema, cost. And every run leaves a record — the exact definitions with their digests, the inputs, the permission plan, the events and the outcome. When something goes wrong, there is something to look at besides a scrolled-away terminal. The second post in this series is about that record.

Try it without a model

The simplest agent is bash: a process's prompt is its script, and each input is an environment variable. The README walks through a "Hello, world" run that needs no model and no API key, and leaves a full record under home/runs/.

Agister also ships a home folder with a multi-lens code review — the workflows I use every day — and an implement process. The third post looks at how that review is composed.

Status

Agister is a working tool I use daily to implement and review code with AI agents. Settings and records are versioned formats (v1alpha1) and may still change. The code is public to read; it is not open source.

Documentation: how a run executes and how settings are read, layered and resolved. Code: github.com/igor1309/agister.